Depth estimation from multispectral satellite imagery: a comparison of conventional and machine learning-based approaches- case study: Kish Island, Persian Gulf

This study investigates underwater depth estimation in the coastal region of Kish Island (KI) in the northern Persian Gulf (PG), Iran, using data from three satellite sensors: WorldView-2 (WV2), Sentinel-2B MSI (S2), and Landsat-8 OLI (L8). Field measurements were conducted with a portable echosounder and a handheld GPS device, resulting in a dataset of 689 points, with 460 used for model training and 229 reserved for validation and accuracy assessment. Atmospheric correction was then performed using the ACOLITE module, due to its adaptability across various sensors. Conventional (linear and ratio transform) and machine learning (ML)-based (random forest [RF], fuzzy inference system [FIS], and adaptive neuro-fuzzy inference system [ANFIS]) techniques were applied to extract depth values from the satellite data. The results showed that L8 provided higher accuracy and precision for both conventional and ML-based methods compared to WV2 and S2. ML-based techniques, especially ANFIS, further improved accuracy, particularly when using WV2. Although depth estimation for areas with less than 2 m of depth was less accurate, both approaches showed high accuracy for depths of 2–4, 4–8, and more than 8 m, respectively, with a mean absolute percentage error (MAPE) ≤ 40%. Final results demonstrated that the ANFIS-L8 model performed best, achieving R² values of 0.951 for training and 0.915 for testing, RMSE values of 0.822 m for training and 1.06 m for testing, and a MAPE of 10.95%. Overall, ML-based methods, particularly ANFIS, demonstrated superior performance, offering valuable insights for coastal and marine applications.